1 /* 2 * Copyright (C) 2018 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 // This test only tests internal APIs, and has dependencies on internal header 18 // files, including NN API HIDL definitions. 19 // It is not part of CTS. 20 21 #include <android-base/file.h> 22 #include <android/sharedmem.h> 23 #include <gtest/gtest.h> 24 25 #include <fstream> 26 #include <string> 27 28 #include "Manager.h" 29 #include "Memory.h" 30 #include "TestMemory.h" 31 #include "TestNeuralNetworksWrapper.h" 32 33 using WrapperCompilation = ::android::nn::test_wrapper::Compilation; 34 using WrapperExecution = ::android::nn::test_wrapper::Execution; 35 using WrapperMemory = ::android::nn::test_wrapper::Memory; 36 using WrapperModel = ::android::nn::test_wrapper::Model; 37 using WrapperOperandType = ::android::nn::test_wrapper::OperandType; 38 using WrapperResult = ::android::nn::test_wrapper::Result; 39 using WrapperType = ::android::nn::test_wrapper::Type; 40 41 namespace { 42 43 // Tests to ensure that various kinds of memory leaks do not occur. 44 // 45 // The fixture checks that no anonymous shared memory regions are leaked by 46 // comparing the count of /dev/ashmem mappings in SetUp and TearDown. This could 47 // break if the test or framework starts lazily instantiating something that 48 // creates a mapping - at that point the way the test works needs to be 49 // reinvestigated. The filename /dev/ashmem is a documented part of the Android 50 // kernel interface (see 51 // https://source.android.com/devices/architecture/kernel/reqs-interfaces). 52 // 53 // (We can also get very unlucky and mask a memory leak by unrelated unmapping 54 // somewhere else. This seems unlikely enough to not deal with.) 55 class MemoryLeakTest : public ::testing::Test { 56 protected: 57 void SetUp() override; 58 void TearDown() override; 59 60 private: 61 size_t GetAshmemMappingsCount(); 62 63 size_t mStartingMapCount = 0; 64 bool mIsCpuOnly; 65 }; 66 SetUp()67 void MemoryLeakTest::SetUp() { 68 mIsCpuOnly = android::nn::DeviceManager::get()->getUseCpuOnly(); 69 mStartingMapCount = GetAshmemMappingsCount(); 70 } 71 TearDown()72 void MemoryLeakTest::TearDown() { 73 android::nn::DeviceManager::get()->setUseCpuOnly(mIsCpuOnly); 74 const size_t endingMapCount = GetAshmemMappingsCount(); 75 ASSERT_EQ(mStartingMapCount, endingMapCount); 76 } 77 GetAshmemMappingsCount()78 size_t MemoryLeakTest::GetAshmemMappingsCount() { 79 std::ifstream mappingsStream("/proc/self/maps"); 80 if (!mappingsStream.good()) { 81 // errno is set by std::ifstream on Linux 82 ADD_FAILURE() << "Failed to open /proc/self/maps: " << std::strerror(errno); 83 return 0; 84 } 85 std::string line; 86 int mapCount = 0; 87 while (std::getline(mappingsStream, line)) { 88 if (line.find("/dev/ashmem") != std::string::npos) { 89 ++mapCount; 90 } 91 } 92 return mapCount; 93 } 94 95 // As well as serving as a functional test for ASharedMemory, also 96 // serves as a regression test for http://b/69685100 "RunTimePoolInfo 97 // leaks shared memory regions". 98 // 99 // TODO: test non-zero offset. TEST_F(MemoryLeakTest,TestASharedMemory)100 TEST_F(MemoryLeakTest, TestASharedMemory) { 101 // Layout where to place matrix2 and matrix3 in the memory we'll allocate. 102 // We have gaps to test that we don't assume contiguity. 103 constexpr uint32_t offsetForMatrix2 = 20; 104 constexpr uint32_t offsetForMatrix3 = offsetForMatrix2 + sizeof(matrix2) + 30; 105 constexpr uint32_t weightsSize = offsetForMatrix3 + sizeof(matrix3) + 60; 106 107 #ifdef __ANDROID__ 108 int weightsFd = ASharedMemory_create("weights", weightsSize); 109 #else // __ANDROID__ 110 TemporaryFile tmpWeightsFile; 111 int weightsFd = tmpWeightsFile.release(); 112 CHECK_EQ(ftruncate(weightsFd, weightsSize), 0); 113 #endif // __ANDROID__ 114 ASSERT_GT(weightsFd, -1); 115 uint8_t* weightsData = 116 (uint8_t*)mmap(nullptr, weightsSize, PROT_READ | PROT_WRITE, MAP_SHARED, weightsFd, 0); 117 ASSERT_NE(weightsData, nullptr); 118 memcpy(weightsData + offsetForMatrix2, matrix2, sizeof(matrix2)); 119 memcpy(weightsData + offsetForMatrix3, matrix3, sizeof(matrix3)); 120 WrapperMemory weights(weightsSize, PROT_READ | PROT_WRITE, weightsFd, 0); 121 ASSERT_TRUE(weights.isValid()); 122 123 WrapperModel model; 124 WrapperOperandType matrixType(WrapperType::TENSOR_FLOAT32, {3, 4}); 125 WrapperOperandType scalarType(WrapperType::INT32, {}); 126 int32_t activation(0); 127 auto a = model.addOperand(&matrixType); 128 auto b = model.addOperand(&matrixType); 129 auto c = model.addOperand(&matrixType); 130 auto d = model.addOperand(&matrixType); 131 auto e = model.addOperand(&matrixType); 132 auto f = model.addOperand(&scalarType); 133 134 model.setOperandValueFromMemory(e, &weights, offsetForMatrix2, sizeof(Matrix3x4)); 135 model.setOperandValueFromMemory(a, &weights, offsetForMatrix3, sizeof(Matrix3x4)); 136 model.setOperandValue(f, &activation, sizeof(activation)); 137 model.addOperation(ANEURALNETWORKS_ADD, {a, c, f}, {b}); 138 model.addOperation(ANEURALNETWORKS_ADD, {b, e, f}, {d}); 139 model.identifyInputsAndOutputs({c}, {d}); 140 ASSERT_TRUE(model.isValid()); 141 model.finish(); 142 143 // Test the two node model. 144 constexpr uint32_t offsetForMatrix1 = 20; 145 constexpr size_t inputSize = offsetForMatrix1 + sizeof(Matrix3x4); 146 #ifdef __ANDROID__ 147 int inputFd = ASharedMemory_create("input", inputSize); 148 #else // __ANDROID__ 149 TemporaryFile tmpInputFile; 150 int inputFd = tmpInputFile.release(); 151 CHECK_EQ(ftruncate(inputFd, inputSize), 0); 152 #endif // __ANDROID__ 153 ASSERT_GT(inputFd, -1); 154 uint8_t* inputData = 155 (uint8_t*)mmap(nullptr, inputSize, PROT_READ | PROT_WRITE, MAP_SHARED, inputFd, 0); 156 ASSERT_NE(inputData, nullptr); 157 memcpy(inputData + offsetForMatrix1, matrix1, sizeof(Matrix3x4)); 158 WrapperMemory input(inputSize, PROT_READ, inputFd, 0); 159 ASSERT_TRUE(input.isValid()); 160 161 constexpr uint32_t offsetForActual = 32; 162 constexpr size_t outputSize = offsetForActual + sizeof(Matrix3x4); 163 #ifdef __ANDROID__ 164 int outputFd = ASharedMemory_create("output", outputSize); 165 #else // __ANDROID__ 166 TemporaryFile tmpOutputFile; 167 int outputFd = tmpOutputFile.release(); 168 CHECK_EQ(ftruncate(outputFd, outputSize), 0); 169 #endif // __ANDROID__ 170 ASSERT_GT(outputFd, -1); 171 uint8_t* outputData = 172 (uint8_t*)mmap(nullptr, outputSize, PROT_READ | PROT_WRITE, MAP_SHARED, outputFd, 0); 173 ASSERT_NE(outputData, nullptr); 174 memset(outputData, 0, outputSize); 175 WrapperMemory actual(outputSize, PROT_READ | PROT_WRITE, outputFd, 0); 176 ASSERT_TRUE(actual.isValid()); 177 178 WrapperCompilation compilation2(&model); 179 ASSERT_EQ(compilation2.finish(), WrapperResult::NO_ERROR); 180 181 WrapperExecution execution2(&compilation2); 182 ASSERT_EQ(execution2.setInputFromMemory(0, &input, offsetForMatrix1, sizeof(Matrix3x4)), 183 WrapperResult::NO_ERROR); 184 ASSERT_EQ(execution2.setOutputFromMemory(0, &actual, offsetForActual, sizeof(Matrix3x4)), 185 WrapperResult::NO_ERROR); 186 ASSERT_EQ(execution2.compute(), WrapperResult::NO_ERROR); 187 ASSERT_EQ( 188 CompareMatrices(expected3, *reinterpret_cast<Matrix3x4*>(outputData + offsetForActual)), 189 0); 190 191 munmap(weightsData, weightsSize); 192 munmap(inputData, inputSize); 193 munmap(outputData, outputSize); 194 close(weightsFd); 195 close(inputFd); 196 close(outputFd); 197 } 198 199 #ifndef NNTEST_ONLY_PUBLIC_API 200 // Regression test for http://b/73663843, conv_2d trying to allocate too much memory. TEST_F(MemoryLeakTest,convTooLarge)201 TEST_F(MemoryLeakTest, convTooLarge) { 202 android::nn::DeviceManager::get()->setUseCpuOnly(true); 203 WrapperModel model; 204 205 // This kernel/input size will make convQuant8 allocate 12 * 13 * 13 * 128 * 92 * 92, which is 206 // just outside of signed int range (0x82F56000) - this will fail due to CPU implementation 207 // limitations 208 WrapperOperandType type3(WrapperType::INT32, {}); 209 WrapperOperandType type2(WrapperType::TENSOR_INT32, {128}, 0.25, 0); 210 WrapperOperandType type0(WrapperType::TENSOR_QUANT8_ASYMM, {12, 104, 104, 128}, 0.5, 0); 211 WrapperOperandType type4(WrapperType::TENSOR_QUANT8_ASYMM, {12, 92, 92, 128}, 1.0, 0); 212 WrapperOperandType type1(WrapperType::TENSOR_QUANT8_ASYMM, {128, 13, 13, 128}, 0.5, 0); 213 214 // Operands 215 auto op1 = model.addOperand(&type0); 216 auto op2 = model.addOperand(&type1); 217 auto op3 = model.addOperand(&type2); 218 auto pad0 = model.addOperand(&type3); 219 auto act = model.addOperand(&type3); 220 auto stride = model.addOperand(&type3); 221 auto op4 = model.addOperand(&type4); 222 223 // Operations 224 uint8_t op2_init[128 * 13 * 13 * 128] = {}; 225 model.setOperandValue(op2, op2_init, sizeof(op2_init)); 226 int32_t op3_init[128] = {}; 227 model.setOperandValue(op3, op3_init, sizeof(op3_init)); 228 int32_t pad0_init[] = {0}; 229 model.setOperandValue(pad0, pad0_init, sizeof(pad0_init)); 230 int32_t act_init[] = {0}; 231 model.setOperandValue(act, act_init, sizeof(act_init)); 232 int32_t stride_init[] = {1}; 233 model.setOperandValue(stride, stride_init, sizeof(stride_init)); 234 model.addOperation(ANEURALNETWORKS_CONV_2D, 235 {op1, op2, op3, pad0, pad0, pad0, pad0, stride, stride, act}, {op4}); 236 237 // Inputs and outputs 238 model.identifyInputsAndOutputs({op1}, {op4}); 239 ASSERT_TRUE(model.isValid()); 240 model.finish(); 241 242 // Compilation 243 WrapperCompilation compilation(&model); 244 ASSERT_EQ(WrapperResult::NO_ERROR, compilation.finish()); 245 WrapperExecution execution(&compilation); 246 247 // Set input and outputs 248 static uint8_t input[12 * 104 * 104 * 128] = {}; 249 ASSERT_EQ(WrapperResult::NO_ERROR, execution.setInput(0, input, sizeof(input))); 250 static uint8_t output[12 * 92 * 92 * 128] = {}; 251 ASSERT_EQ(WrapperResult::NO_ERROR, execution.setOutput(0, output, sizeof(output))); 252 253 // This shouldn't segfault 254 WrapperResult r = execution.compute(); 255 256 ASSERT_EQ(WrapperResult::OP_FAILED, r); 257 } 258 #endif // NNTEST_ONLY_PUBLIC_API 259 260 } // end namespace 261